Frequency Attention Network Accurately Detects AI-Generated Images Across Visual Benchmarks

Computer vision researchers have developed an advanced dual-domain neural framework that detects artificial intelligence-generated images with detection accuracies exceeding ninety-six percent across diverse visual datasets. As modern generative models including generative adversarial networks and diffusion systems produce photorealistic synthetic visuals, differentiating computer-generated content from authentic photography has become an urgent priority for digital media verification and visual forensics. Traditional detection systems relying purely on spatial convolutional networks often overlook subtle spectral anomalies, while frequency-only detectors frequently overfit to generator training artifacts. To resolve these operational vulnerabilities, investigators designed the Frequency Attention Taylor-aware Neighbor Baikal Network, known as FTNB-Net, which modifies the learning rules of the Frequency Attention Squeeze Network through an advanced composite loss function. Evaluated across standardized image collections, the architecture demonstrated a recall of 97.479 percent, a precision of 96.611 percent, and an overall classification accuracy of 96.848 percent on specialized fruit imagery, while achieving 97.368 percent accuracy on broader synthetic artistic collections.

The foundational structure of the framework builds directly upon the Frequency Attention Squeeze Network, an architecture engineered to process visual inputs through coordinated stages of standardized data preprocessing, image augmentation, frequency-domain spectral analysis, dual-branch feature mining, and probabilistic classification. The detection procedure commences by standardizing incoming image dimensions to a uniform resolution of 224 by 224 pixels, ensuring operational compatibility across deep convolutional layers and transformer processing blocks. Following spatial resizing, pixel intensity values are normalized into the continuous mathematical range of zero to one, stabilizing internal weight updates during optimization and facilitating faster numerical convergence. To ensure that the resulting neural model remains robust against incidental perturbations and real-world modifications, the normalized visual data undergo systematic image augmentation routines. These transformations encompass random horizontal and vertical flipping operations alongside color jittering adjustments that alter brightness, contrast, and saturation, training network layers to learn persistent discriminative features rather than memorizing superficial color distributions or directional orientations.

Following preliminary data preparation, the pipeline performs detailed frequency-domain analysis to capture spectral anomalies characteristic of generative synthesis engines that remain imperceptible within conventional spatial representations. Generative algorithms typically introduce unusual high-frequency energy distributions or unnatural periodic regularities, contrasting sharply with the smooth, continuous spectrum exhibited by natural photography. To extract these latent signatures, the framework applies a two-dimensional Fast Fourier Transform to the augmented image matrices, mapping spatial pixel coordinates into frequency coordinates and isolating structural Fourier magnitude representations. After targeted frequency filtering isolates anomalous high-frequency bands indicative of artificial generation, the system applies an Inverse Fast Fourier Transform reconstruction to convert the filtered spectral signals back into the spatial domain. This inverse transformation retains critical frequency characteristics while yielding a reconstructed spatial image matrix that emphasizes subtle discrepancies between real and synthetic items, providing an enriched foundation for subsequent multi-level feature extraction stages.

Feature extraction within the framework operates across two complementary structural streams to capture both broad contextual structures and microscopic visual artifacts from the reconstructed spatial representations. Global feature extraction is managed by a Swin Transformer Encoder, which models long-range contextual relationships and holistic semantic information across the entire visual field. The encoder divides the input feature map into non-overlapping spatial patches of four by four pixels, generating a total of 3,136 distinct visual patches arranged across a grid of fifty-six by fifty-six elements. These localized patch vectors are projected into higher-dimensional feature spaces and processed through shifted window-based self-attention mechanisms and regular window multi-head self-attention modules. By confining self-attention calculations to local windows while permitting cross-window interactions through shifted partitioning between successive layers, the transformer reduces theoretical computational complexity from quadratic to linear relative to image dimensions, enabling the model to learn hierarchical, multi-scale contextual dependencies efficiently.

Operating simultaneously alongside the global transformer stream, a specialized local feature extraction branch captures micro-pattern variations, edge configurations, and structural textures that are frequently corrupted during artificial synthesis. This pathway employs a Convolutional Neural Network backbone designed to derive localized convolutional feature maps through parameterized filter kernels and nonlinear activation functions. To maximize the discriminative capability of these localized representations, the convolutional layers are integrated with Squeeze-and-Excitation attention modules that recalibrate channel dependencies dynamically. During the squeeze stage, a global average pooling operation condenses spatial dimensions across each channel, establishing a compact descriptor that reflects global contextual importance. In the subsequent excitation phase, this channel descriptor passes through a bottleneck network consisting of two fully connected layers with non-linear activations to compute channel-specific attention weights. These weights scale the original convolutional feature maps through channel-wise multiplication, accentuating informative textural cues while dampening redundant signals before forwarding features to a final SoftMax binary classifier.

A central mathematical contribution of the FTNB-Net framework is the formulation of the Taylor-aware Neighbor Baikal Loss, an optimization function established to refine the learning rules of the network during training. Conventional objective functions commonly focus strictly on minimizing binary misclassification errors, frequently failing to maintain training stability when visual boundaries between authentic and synthetic images are exceedingly fine. The Neighbor loss component acts as an even, regression-based loss function derived from a Log-cosh formulation, calculating continuous mathematical proximity between network predictions and true class targets while ensuring balanced gradient updates and curbing extreme numerical deviations. Complementing this behavior, the Baikal loss operates as a probabilistic objective derived from cross-entropy formulations, measuring class probability discrepancies to widen margin separability between authentic and artificial classes. To harmoniously merge these divergent loss functions without introducing gradient instability, the architecture utilizes a second-order Taylor series expansion based on functional derivatives, creating a smooth, composite objective that guides stable weight convergence.

The resulting hybrid loss formulation combines the Taylor-expanded Neighbor loss and scaled Baikal loss into an integrated optimization target whose parameters are updated through decomposed learning rules. Using standard Stochastic Gradient Descent principles, the researchers performed partial differentiation with respect to intermediate network weights to examine the individual mathematical contributions of both loss components during optimization. This formulation guarantees that as the perturbation parameter approaches zero, the resulting gradient updates systematically balance error penalization and derivative approximations across each training epoch. By coordinating the derivative terms derived from the second-order Taylor expansion across sequential iterations, the decomposed update rule prevents abrupt gradient oscillations and facilitates smooth optimization trajectories. Consequently, the network tunes its internal convolutional filters and attention parameters to isolate delicate manipulation residues that baseline cross-entropy loss functions consistently overlook, substantially improving feature discrimination between authentic and artificially generated visual content.

The experimental validation of the FTNB-Net architecture was carried out using two distinct image repositories designed to benchmark synthetic image forensics under rigorous conditions. The first collection, designated as Dataset-1, comprises 306 high-resolution images of apples occupying approximately 86.06 megabytes of storage, evenly divided between 150 artificially generated specimens and 156 authentic photographs across green and red fruit varieties. Following standard augmentation protocols, this initial fruit collection was expanded to 1,530 samples to ensure robust training exposure. The second repository, designated as Dataset-2, represents a substantially larger benchmark containing 21,600 images totaling 14.44 gigabytes, created specifically to separate human-made imagery from artistic synthetic content produced by generative engines including MidJourney and DALL-E. For experimental evaluations, each dataset was partitioned using a strict distribution ratio allocating ninety percent of available samples to model training, five percent to validation monitoring, and the remaining five percent to formal testing.

Performance evaluations revealed that the FTNB-Net model achieved superior detection scores across both testing environments, particularly when configured with smaller processing batch sizes during optimization. When trained on Dataset-1 with ninety percent training data and a batch size of sixteen, the network recorded a classification accuracy of 96.848 percent, accompanied by a recall of 97.479 percent and a precision of 96.611 percent, outperforming larger batch configurations of 32, 64, and 128 samples. On Dataset-2, the network registered an accuracy of 97.368 percent, a recall of 97.688 percent, and a precision of 96.888 percent under identical training conditions. Receiver Operating Characteristic curve evaluations further confirmed high discriminative reliability, with FTNB-Net securing a True Positive Rate of 99.456 percent on Dataset-1 and 99.998 percent on Dataset-2 at a False Positive Rate threshold of eighty percent. In comparative noise robustness trials against eight baseline detection frameworks, FTNB-Net maintained 92.999 percent accuracy on Dataset-1 under twenty percent noise, while also sustaining 85.989 percent accuracy during cross-dataset evaluation where the system was trained on Dataset-1 and evaluated on Dataset-2.

Model interpretability was rigorously confirmed through Gradient-weighted Class Activation Mapping visualizations, which demonstrated that FTNB-Net effectively focused on discriminative frequency and textural anomalies rather than background artifacts. Systematic ablation experiments proved the individual necessity of each component, as removing the Taylor Neighbor Baikal Loss, Swin Transformer Encoder, convolutional blocks, or preprocessing lowered baseline accuracy from 96.848 percent down to 90.467 percent on Dataset-1. In real-world degradation assessments encompassing JPEG compression, spatial blur, color space conversions, and image resizing, the framework consistently outperformed established models including MEXFIC, DiffCoR, APN, and Vision Transformers. Under adversarial evasion, poisoning, and inversion attacks, FTNB-Net sustained high detection rates exceeding ninety percent, while ten-fold cross-validation paired t-tests confirmed that all improvements were statistically significant with p-values below 0.05. Finally, computational profiling recorded rapid inference times of 0.18 to 0.20 seconds per image, memory consumption below ninety-nine megabytes, and processing demands of three to five GFLOPs, establishing a scalable and efficient solution for practical multimedia verification.

Subject of Research: Detection of AI-generated synthetic images using frequency analysis and Taylor Neighbor Baikal loss

Article Title: A frequency attention squeeze framework enhanced with neighbor Baikal Taylor loss for AI-generated image detection

Article References: Bairam, M., Velpula, R., & Naik, R. L. (2026). A frequency attention squeeze framework enhanced with neighbor Baikal Taylor loss for AI-generated image detection. Discover Computing, 29(1), Article 800. https://doi.org/10.1007/s10791-026-10657-1

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10657-1

Keywords: FTNB-Net, AI-Generated Image Detection, Frequency Analysis, Fast Fourier Transform, Inverse Fast Fourier Transform, Taylor Neighbor Baikal Loss, Swin Transformer, Squeeze-and-Excitation, Convolutional Neural Networks, FAS-Net, Deepfake Forensics, Spectral Feature Extraction, Adversarial Robustness, Grad-CAM Explainability, Synthetic Media Detection, Image Forensics, Computer Vision